For example:
Whenever I'm dealing with a data processing pipeline I tend to design things as if they would be used from a unix command line. I tell the team to make their classes etc as if they would be wrapped with a simple chunk of code to parse a text stream then serialize to it later. During the course of testing, I almost always will then write those tools to allow simple command line pipes to push test data around. It helps prevent reliance on certain toolsets and certainly helps stop us from conflating bugs in our code with bugs or operator error about a pipeline framework. Later we can go back and more closely integrate with the chosen operating environment, if performance demands it. On the other hand going from a situation where code is written to the environment and migrating to another environment can be very daunting, difficult and error prone.
So seriously? You almost never need to read & write to streams?
Don't tailor from-memory standard library questions to bits of code that almost nobody ever has to use.